Compare document parsers for RAG
Load a document, add the text another parser produced from it, and see which reading lets your questions find the right passage. Free, in your browser. Nothing is uploaded.
- 1Load your document as Reading A, and the other parser's .txt or .md output as Reading B
- 2Add questions with the words the right chunk must contain
- 3Press Compare and read which reading finds more answers
About the example: a made-up pay table. In A, each person's name appears only on the first of their three rows, as a merged cell often comes out of a parser. In B, every row carries the name. The questions name a person and ask for a figure from one row.
Reading A: your document
Reading B: the other parser's output
Questions
For each question, the words the right chunk must contain, for example 398,000. Separate answers in different chunks with ; and phrases that must be in the same chunk with +.
Settings, the same for both
Result
What it is for
- See whether one parser keeps the information your questions need, such as table rows with their names, headings and numbers.
- Compare Docling, docparser.net, LlamaParse, Unstructured, a PDF library or any other parser on your own document, with the same chunking and search.
- Find out what each parser did to the same passage, side by side, before you build a pipeline on it.
- Check that switching to a cheaper parser does not cost you answers.
How it works
Parsers are usually compared on how closely the text matches the page. For RAG the question is different: when someone asks something, does the right chunk come back? A table whose rows lose their names, or a heading that is dropped, can look fine to the eye and still make retrieval fail.
This page takes two readings of the same document, cuts both into chunks with the same settings, and runs your questions against each. Reading A is your document as TryRAG reads it: upload the PDF, or paste its text. Reading B is what another parser produced from the same file, which you bring as a .txt or .md file or paste in. For every question it shows where the right chunk was found in each reading, which reading did better, and the chunk each one returned.
It does not run the other parser for you: parse the file with it, save the output as text or Markdown, then load it as Reading B. Everything is worked out in your browser and nothing is uploaded. With fewer than about 30 questions, small differences may be chance, so treat a result as a pointer and check it on more questions.
Questions
How do I compare two parsers?
Upload your PDF as Reading A (or paste its text). Parse the same file with another parser, such as Docling or docparser.net, and save the result as a .txt or .md file. Load that as Reading B, name both, add questions with the words the right chunk must contain, and press Compare the readings.
Which parsers can I compare?
Any parser whose output you can save as text or Markdown: Docling, docparser.net, LlamaParse, Unstructured, PyMuPDF or a cloud service. Reading B takes only .txt and .md files, or pasted text. Reading A can be the PDF itself, which TryRAG reads in your browser, or text.
What do I type in "must contain"?
A few words that the right chunk holds, for example 398,000. Separate answers in different chunks with ; and phrases that must be in the same chunk with +, for example Dana Ortiz + 398,000.
Why can two readings of the same page score differently?
The chunks are cut from the text, so what the parser keeps decides what a chunk says. If a table row has lost the name it belongs to, or a heading is missing, the chunk no longer matches a question that names them. The same settings are used on both sides, so the difference comes from the text.
Is this a benchmark of parsers?
No. It tells you which parser works better for your document and your questions, not which is best in general. Results on a handful of questions can be chance.
What is the difference from RAG Battle?
RAG Battle compares settings, document versions or prompts in general. This page is for comparing parsers: Reading A is your document as TryRAG reads it, Reading B is the output you bring from another parser, and it shows the chunk each one returned for every question.
Is my text uploaded?
No. Everything runs in your browser tab. Your documents and questions never leave your device.